—

Estimating Noise in W Systems

Jtijeet Roycho\vdhury Aper Detir

~aJee~tipdemir} @research.bell-labs.com
BeH Laborat~ties
Murray Hall, Netv Jersey, USA

Abstract periodic signrd, i.e., discrete tones in the spectrum, which is required
Prdlcting noise performance of W systems is much more mmplex thm for in an ided mixing fi.e., frequency translation) operation. Any noise
kear ones. me common practice of retrofitting finw noise concepts often in the device (i.e., frequency synthesizer) that generates the LO signrd
Id to wrong answers. For a proper understanding of M noise, nosdirrtiti~ restits in corruption of these discrete tones. me way a corrupted LO
and synchronition, i.e. phase noise in osdatom, have to be accormtd for signrds tiects the information signal in the mixing operation is quite
correctiy, using stochastic process=. h this tutond, we outiie how to do so different than, for instance, the noise in an amplifier affuts it.
for non+sci~atory and ostilatory M systems. \Ve emphasim basic con~pts
and important results, wtie providing sufficientdetails to maintain rigour. Our me origin of the information signal is, most of the time, digitd,
presentation W cover i.e., a bh stream. me effect of the notiinearities and noise in the N
● basic stochastic concepts for noise anrdysis
amdog front-end manifests itself as bit errors one detects by compar-
ing the demodulated information signal bit stream at the output of the
● notrstatiosrarity,cyclostationadty and frequency-correlationof noise
receiver with the ones that was used to modulate the carrier at the trans-
● mixing noke in driven (nonmsdatory) systems mitter. me ultimate performance measure for an ~ transceiver is the
. pbme noise in osc~ators bit-error-rate @ER). One would Eke to analyze and simulate the whole
● computatiomdtechniques N transceiver system before fabricating the WC to make sure that it
OurgoA is to enable designers and CD engineemto understandthe basic flow mmts the BER specification in a varie~ of circumstances. If it does
of W noise amdysisand the current Htemture. not m~t the specification, one wordd tike to identi~ the dominant
sources of error so that the problem can be fixd by focusing on that
1 htroduction part of the system and not by redesigning the whole thing. If it meets
me increasing demand for low cost, high performance, hi@y inte- the specification with a margin larger than requird, one would Eke to
grated solutions for RF mobile conummications products, and time-to- know, design of what parts of the system cart be relaxd so that the
market pressures, made the currentiy required number of design iter- specifications are still met with lower cost, lower power dissipation,
ations for RF ICS unacceptable. me analog front-end continues to be etc. ~us, one wodd like to have ways of characterizing the individud
the bottienwk in the design of RF communications transceivers. performances of the components that make up the system and know
me basic finctionrdity of the ~ anrdog front-end of a commu- how they contribute to the overall system performance. Moreover, one
nications r-eiver is to extract the low-bandwidth information signal wotid like to use the andysis/simdation techniques not ody for veri-
from a high frequency carrier signrd which rides on top of it by modu- @ing the performance of an rdrertdydesigned system or subsystem but
lating its arnp~tude, phase or frquency. Even though there are many rdso in making design decisions.
different radio architectures, dl of them consist of several stages of In this tutorial, we present an overview of the mathematicrd con-
arnphfication, filtering and mixing operations. Ml of these operations cepts from the theory of dynamical systems and stocktic processes
act on the information bearing signfl, on top of a carrier or at base- that are used in developing an anrdysis/simulation methodology for the
band, and modi~ some of its characteristics. mere are dso intemrdly so-cold RF noise problem. Even though we will touch upon the ac-
generated signals, e.g., a clock signrd for a switched-capacitor circuit tual (numerical) dgonthms used for analyzing ~ noise, our empha-
or the output of a frequency synthesizer, which are requird in one of sis wifl be on the concepts used in the mathematicrd representation of
the amplification, filtering or mixing operations. the signals and systems, and how input signal representations cart be
Verifying the correctness of the basic ided functiontity of a given prosed through the system representations for the output sigmd repre-
~ front-end is trivial, because of the extremely simple nature of the sentations.
operation, urdike, for instance, the operation of a control logic cir-
We wiU first concentrate on the signal representation problem. me
cuitry. What is far from trivird is to verify that the basic functionrdity
modeting of noise-like signals using deterministic and stochastic rep-
is corrupted below a specified acceptable level by the non-idd oper-
resentations will be discussed and compard against each other from
ation of the system components caused by the norr-lirrean.tiesin the
the viewpoints of geneflhy, efficiency and practicality. Both time
M signal path and noise. Non-finearities in the signal path corrupt the
and frequency domain concepts in stochastic modeling will be cov-
information bearing signal by distorting it. Non-finearities rdso crate
ered. We will introduce and illustrate the concepts of stationarity,
corruption by translating other undesird information signals (which
non-stationari~, cyclo-stationarity, probability density functions, and
are in the spectrum raeived) in frequency and adding them to the de-
correlah.on both in time and frequency domain, and discuss how they
sired signal. Noise, which is generated either within the components
arise in dting with the RF noise problem. me issues that arise in
of the devices that make up the analog-front end (i.e., therrnrd, shot,
modeling the noise generated within the devices that make up the N
flicker noise), or in other parts of the IC (e.g., digitd switching noise
front-end due to the furtdamentrds of the IC device operation (thermal,
coupled through the power supply fines, the chip interconnect and the
shot, flicker noise, etc.), which is unavoidable, as well as the ones that
IC substrate), corrupts by coupling to the signal path and adding on
are generated in other parts of the chip (digitrd switching noise coupled
top of the information signal. Apart from direcfly adding on top of the
througfi the substrate, supply lines and the interconnect), which has a
information signrd in the sigmd path, noise dso causes distortion and
interference from other signals in the s~tmm. ~s happens through significant impact on the performance of the W components, will be
brought out. me importance of the Gaussian density, white noise (no
the local oscillator LO) signal that is used in the mixing operation,
correlation between time samples) and its dud, stationary noise (no
which is not in the sigrtrd path. me LO signrd, idedy, is a perfecfly
correlation between frequency samples), will be explained.
Petission to makedgiti or tid copies of aU or pti of h \vorkfor ~ti or Next, we consider the system representation problem, i.e., the mod-
&moom me k ~ted without fm protided that copi= are not made or &&b efing of the components that makeup the front-end using partifl differ-
utti for profit or comerti ad~,mrtageand that copies bear this notice and tie fuU
citation on We fit page. To copy othmtie, to repubkh, to past on semm or to ential quations &DEs), ordinary differential equations (ODES), trfi~s-
r~ti%ute to h~, rquires prior -c ~tiion and/or a f=. fer functiorls, etc. me implications of a system being non-autonomous
ICW98, Sm Jose, CL USA
0199SA~f 1-58113@S-ti9S/Ml l.S5.W

199
(e.g., arnptifiers, mixers), uutanomous (e.g., oscillators), or “semi”- of each member of the ensemble x(t) (or its square for the power), to
autonomous (e.g., phase-locked loops, frequency dividers) from the obtain a new ensemble X(f), parametrized by the frequency ~. Such
noise analysis viewpoint will be discussed. The distinction between a definition is not, however, strictiy correct, simply because members
how noise ti’ects a non-autonomous system versus an autonomous one of the x(t) ensemble may not have finite energy and hence cannot be
will be emphasized. In particular, the so-crdled pke noise problem Fourier-transformed. Nevertheless, a technic~ly correct definition that
that is associated with oscillators (an autonomous system) will be cov- retains this intuition is possible. The transformed ensemble X(f) is a
ered. The concepts of a system being linear, nordine~, time-i~variant stochmtic process in frequency, with mean px(~) and variance ~~(f)
and time-varying will be explained from the perspective of noise ~~- Providng info~ation about its averages. We will not follow the appar-
ysis. enfly intuitive course above to understand stochastic processes in the
Given the mathematicrd representations of the desired siW~s, frequency domtinl, but we will return to it when necessary. Instead,
noise and the system they are influencing, we would like to charac- we wi~ continue by considering the autocorrelation fiction of x(t),
terize thesystem output using a similar mathematical rePr~entatiOn a concept of cen~~ impofi~ce in noise analysis, both from the time-
as the one used for the inputs. These characterizations should be such and frequency-domain points of view.
that various figures of merit (e.g., noise figure, single-sideband phase
The autocorrelation fiction R=(t,z)is the expected value of
noise, timing jitter, etc.) RF circuit and system designers are used
to describe the performance of their circuits can be derivdcomputed x(t)x(t + 7), i.e., the quantity x(t)x(t + T) averaged over dl members
easily. There are two, seemingly contradictory, requirements on noise of the ensemble of x(t). The function denotes the correlation between
vrdues of the stochastic process at different times. If R~(t, z) depends
analysis techniques They should be rigorous, but at the same time
provide intuitive characterizations that can guide design. Most of the ordy on z, i.e., otiy on the time difference between the two timepoints,
currently available rigorous noise analysis t=hniques are not widely x(t) is cdld a stationa~ stocktic proces$ if there is a dependence
used by designers because they fail to provide intuition. on t aswell, it is crdled nonsfationa~. Stationary processes are of
great importance in circuits and systems. We will first consider sta-
RF systems are complex and large, in the sense of the number of .
v~ables, ~fferenti~ equations and the noise SOUrC= that Me USd to bonW ‘ro~=sm~
model the dynamics. This puts a restriction on the kinds of noise andy- If x(t) IS stationary, the Fourier transform of its autocorrelation
sis dgonthms that can be usd for practicrd circuits. Usually, the more R=(z is cdld the power spectral density of x(t), and denoted by
gened the noise analysis algorithm is, the less efficient it is. To derd ~~(j 1. The pow:r spwtrfl density (or PSD) is a useful characteriza-
with the complexity and the large size of the noise anflysis problem, tion of a stochastic process, for it can be measured directiy on a spec-
one uses the same principle used in the design of these systems, i.e., ex- trum amdyzer. This is so because there is a direct connection between
ploit the hierarchy and handle the components that makeup the system the PSD and the Fourier transforms of the squares of the stochasic pro-
separately. However, in design, one still has to put the components to- cess ensemble denoted by X(j) above. It can be shown that the mean
gether and make sure they work together. Same in noise analysis. One value of the Fourier transforms of the squared process, at a given vrdue
can design a mixer and verify that it meets a certain noise figure specifi- of frequency f, is equrd to Sm(f). Further, two random variables X(fl )
cation, and design an oscillator and verify its phase noise performance. and X(f2) can be shown to be uncorrelated if fl # f2. Spectrum an-
Then they need to be put together, and verify that the corruption of the dyzers display Fourier transforms of the squared input, averaged over
information signal @ecause of the non-idd function of the mixer and separated sections of time, thus approximating an ensemble average
the oscillator) that is frequency translated is at acceptable levels. We and measuring the power specti density directiy. Another property
will present an overview of noise anrdysis ~goritbms that have bwn, of the PSD is that its inte@ over frequency equrds the variance of
or are &lng, developed. We will cover linear time-invariant anrdysis x(t). This is the toti noise power, an important figure of merit.
with stationary noise, finear (periodicdly) time-varying anrdysis with If the PSD is independent of frequency, i.e., S~(f ) is a constant,
(cycle-stationary) non-stationary noise, phase noisdtimirrg jitter and- the process is known as white no~se.This corresponds to the autocome-
ysis and nodinear noise anrdysis with large noise signrds. Rather than lation function Rn(t) Ming a delta function in c, implying that neigh-
the details of the formulation of these analysis ~gorithms, our empha- bouring points of the process are totally uncorreleatd. Intuitively, this
sis will be on their gened features and how they fit into a hierarchi- suggests underlying physicrd processes that are extremely rapid or with
crd methodology for anrdysis and design of RF front-ends. We will very short memory. Many importmt sources of noise in circuits can be
dso discuss reduced-order and black-box modeling t~ti~u~, both modelled adequately as white noise. Thermal noise in resistors and
for the noise signrds and the systems, which are going to be mdsspens- shot noise in semiconductors are white, provided bias conditions are
able tools in derding with both the complexity and the moddmty of steady.
RF circuit and system dwign. me concepts introduced in the tuto- The concepts of stationarity suffices for noise cdcrdations in sys-
rial will be illustrated with noise anrdysis examples for RF circuits and tems in a smrdl Enear region about a DC operating point, such as finear
systems. amufifiers. However, some comrronents in RF systems, such as rnix-
2 Represenhtion of noise si~~ ers; do not operate in’s smd regi”onaround a qui&cent point, but have
Iarge-signrd swings that are crucial to their operation. Stationary con-
htuitively, noise is often visu~lsed as an undwirable, fuzzy wave- cepts no longer suffice for d~cribing noise in such systems. It is easy
form corrupting a signal on, say, an oscilloscope screen or spectrum to appreciate why. Consider a switch that turns on and off periodicdy,
anrdyser. Implicit in this intuitive picture is the important notion that either passing its input through or blocking it completely. This iderd
a noise waveform is not known exacfly, but ordy in terms of average switching mixer cannot be time-invariant, since the periodic switch
spread. The theory of stochastic processes enables us to concretize control makes the output dependent on the time in the cycle that an
this notion and use it to amdyse complex systems. A full review of input is apphd. Furthermore, if the input is a station~ stochastic
stochastic processes is beyond the scope of this tutori~, here, we re- process, the output is no longer stationary, for its power is zero when
mind the reader of some the impofiant concepts relevant to our further the switch is off, whereas it is the same as that of the input when the
discussion. switch is on. Since the output power vari= with time, the output noise
x(t) is called a stochastic processor random process if for any fixed is not stationary.
vahre of time t,x(t) is a random variable. x(t) therefore represents an
If x(t) is nonstationary, its autocorrelation function R=(t,z)is a
entire collection or ensemble of wavefom. Like any random variable,
function oft. When the dependence on tis periodic or quasiperiodic,
x(t)has a mean Kx(t) and variance a:(t). While it is meaningless to the process is crdled cyclostationq. Cyclostationary processes usually
ask what the vahre of x(t) is at a given time t(forx(t) represents a col- arise in systems such as mixers, that are in periodic or quasiperiodic
lection of waveforms, which are not individually known), the mean and steady state.
variance provide useful imforrnation about averagm of the ensemble.
The autocorrelation function of a cyclostationary process can be
In communication systems, one is often interested in frquency-
domain representations, i.e., Fourier transforms of time-domain wave-
forms. It may be asked what the Fourier transform of a stochastic l~ou& pssible, MSms outto k comidembtymominvolvedtim he dtemtive
process x(t) is. In principle, one may consider the Fourier transform pmenti

200
expanded in a Fourier series in t: which is a generrdization of (3) to L~ systems. For an arbitrary input
with X(f) = F {x(t)}, the output is
(1)
X(x)(t) = ~m H(t,f)X(f)exp(j2zft)df (9)
-m
where Ri(z) are called harmonic autocorrelation functions. Typicrdly, a
finite number of harmonics are sufficient to describe the process. The A linear system is (linear) periodicrdly time-varying (L~), if the
Fourier transform of Ri(~), denoted by Si(~), is called the harmonic impulse response is periodic in t:
power spectd density or HPSD. JVe observe that while a stationary
process has a single autocorrelation function and PSD of one argument, h(r+T,t)=h(t+T+ T,t+T) (10)
a cyclostatiorrary one has many of them.
If r(t) is cyclostationary, it can further be shown that the Fourier with Fourier series representations for the impulse response and the
transformed process X(f), dhrded to previously, is no longer uncor- transfer function
related at different values off. In fact, X(f) and X(f + ifo) can be
shown to be correlated with value Si(f ). This phenomenon is known h(t +T,t)= ~ hn(z)exp (j2mfCt) (11)
as frequency correlation, and is equivrdent to that of having nontrivird n=--
HPSDS. The stationary part of the power spectrum (i.e., the compo-
nent that is independent oft) is given by So(f) and denotes the aver-
age noise power. This is usually the ofly output quantity of inter~t.
However, as explained in Section 4.2, it is very important to keep track H(t, f) = ~ Hn(f +nfc) exp (j2mfCt) (12)
of the other HPSDS during noise analysis of time-varying systems, for n=-m
they can affect SO(f) at the outputs.
where Hn(f ) = F {hn(T)} are the harmonic transfer functions. If the
3 System representation for noise mdysis input to an L~ system fi is x(t) = exp (j2zft), then the output is
Electronic circuits as dynamicrd systems are modeld with partird
and ordinary differenti~ equations, transfer functions, finite-state ma-
chines, etc. For the M noise anrdysis problem, a system of differ- ~(x)(t) = ~ Hn(f +nfc)exp(j2n(f +nfc)t) (13)
enti~dgebraic equations and transfer functions are most appropriate. n=-=
Transfer functions are especially use~, because they represent the
which is a spwid case of (S). The Fourier transform of the output of
system components in frequency domain, the domain of choice for W
design, and as the basis for input-output black-box and reduced-order an L~ system is Y(f) = ~~=_=Hn (f )X(f + nfc), where X(f ) is
models. the Fourier transform of the input. This is a gener~ization of (6) to
Let us define a system as a mapping y = ~(x) from from an in- L~ systems.
If a sin@e complex exponentird at frquency f is input to an L~
put x(t)into an output y(t). A system ~ is said to be linear if system, the output is dso a single complex exponential at frequency
ti(axl + bx2) = afi(xl) + bfi(x2), and time-invariant if fi(x(t + f with a scrded “amplitude”, where the scrding is set by the trans-
T)) = X(x) (t + T). For a linear system, the impulse response is given fer function H(f). For an LN system, the output for a sin~e com-
by /z(t,
u) = ~(S(t – u)). For an arbitrary input, the system output is plex exponentird input, in gened, contains a continuum of frequen-
given by ci~. For L~ systems, from (13), we observe that the output cor-
rwponding to a single complex exponential at frequency f is a sum-
X(x)(t) =/- X(u)h(t,u)du (2) mation of complex exponentids at frequencies f+ nfc, where fc is
-- the fundamenti frequency. An L~ system can be used to model
a mixer as a single-input ~ single-output ~ system considering
If the system is time-invariant, then h(t,u) = h(t – u). If the input the LO as part of the mixer itself. Such a model captures frequency
to an L~ system is a complex exponential at frequency f, x(t) = translation which is the basic function~ity of a mixer. On the other
exp (j2nft), then the output is hand, an L~ model cannot capture nordinearities of the signal path.
The system transfer fiction concept can be generalized to nonlinear
~(x)(t) = H(f) exp(j2nft) (3) time-invariant or (penodicrdly) time-vqing systems to capture sig-
nal path noflinearities. For a nordinear periodically time-varying
where H(f) is the Fourier transform of the impulse response h(t),
system, the output corresponding to a single complex exponentird at
frequency f is a summation of complex exponentirds at frequencies
H(f) =/- h(r – u) exp (–j2zf (t – u))du (4) kf+nfc, k= 1,.. .,~, n~—~ ,... ,~, where f= is the fundamental
-m
frequency.
and is called the system tr~erfiction. For an arbitr~ input with 4 Noise mdysis
Fourier transform X(f) = F {x(t)}, the output is
4.1 Linear time-invariant noise analysis tith stationa~ noise
The propagation of stationary noise in L~ systems is relatively simple
= /- H(f)X(f)
fi(.r)(t) exp (j2nft)df (5) to amdyse and to understand intuitively. This is because the PSD of
-m
a stationary signrd behaves similarly to a deterministic sign~ when it
with the Fourier transform passes through a L~ transfer function. The power spectral densities
of the input and output processes x(t) and y(t) of a L~ system with
y(f) = F {fi(X)(t]} = H(f) X(f). (6) transfer function H(f) are related by:
By anrdogy with (4), the system transfer function H(t,f) for a linear (14)
time-varying (LW system is defined by ~}y(f) = l~(f)12~H(f)
L~ noise computations dl essentially implement the above rela-
H(t, f) = / h(t, u) exp (–j2nf (t – u))du (m tion. For a large dynamic system with many outputs and inputs, a
-- matrix version of (14) is usually computed efficiently using adjoint
Note that, in contrast to H(f) in (4), H(t, f) in (7) is a finction of both techniques.
f and t.If the input to an LN system fi is x(t) = exp (j2nft), then 4.2 Linear @eriodicrdly) time-varying noise analysis tith (cycl~
the output is sbtionary) non-stationary noise
Analysis of nonstationary noise in a LN system requires the machin-
X(x)(t) = H(t, f ) exp (j2zft) (s) ery of stochastic differential equations, which we will not discuss here.

201
me special case of cyclostationq processes in a LPTV system, how- power spectrrd density. For the perfect periodic signal x.(t), the power
ever, is captured by a relation similar to (14). The HPSDS of the input spectrrd density has S functions located at discrete frequencies fi.e.,
process r(r) and the output process y(t) are related by the harmonics). The phase deviation u(t)spreads the power in these S
functions in some specific forms depending on the stochastic properties
~,?(f) = wfl~n(f)~ (f)> (15)
of the noise sources. The exact shape and level of spectd spreading
where S(f) is now a multi of HPSDs arranged in a Toeplitz-like struc- for an LO sign~ is extremely important from a practical point of view,
ture, and ~~) is a similar matrix of harmonic transfer finctions of the since it is the direct cause of undesired phenomena such as interchan-
L~ system. The computation of (15) can be achieved efficiently, nel interference in the mixing operation.
even for large systems. 4.4 Notilnear noise analysis with large noise signrds
\Ve have already observed that the PSD of stationary noise is acted me linear noise analysis techniques we outlined in Section 4.1 and
upon by a L~ system much as a deterministic signal would be. The Section 4.2 are based on a perturbation or smrdl-sigmd amdysis assum-
situation is similar for cyclostationary noise through an LPTV system. ing that both the noise signrds and the deviation of the response of the
Just as the harmonics of a deterministic signal get mixed or frequency- circuit due to the noise signals are small. In Section 4.3, we argued
translated by passage through a LPTV system, so do the HPSDS of that, for oscillators, even when the noise sigmds are assumed to be
cyclostationary noise. And just as in the deterministic case, the DC smtil, the response deviation is not. However, by exploiting the unique
vahse of the output can depend on the harmonics at the input, so can nature of oscillators, we were able to go around this difficulty and de-
the station~ HPSD at the output depend on dl the HPSDS at the input. vise a rigorous and efficient noise amdysis technique. In other appli-
This is why the HPSDS need to be kept track of during anrdysis, even cations, the “small noise” or “small response deviation” may not be
if it is only the stationary component at the output that is of interest. justified. For instance, the information signal itself in CDMA wireless
4S Phase noise and timing jitter anrdysis communications systems is noise-like, and b~t modeled as a stochas-
Oscillators constitute a special class among noisy physicrd systems: tic process.
their autonomom nature makes them unique in their response to per- One approach in derding with the full nosdinear noise analysis
turbations. Introducing even small noise into an oscillator leads to problem is the so-called Monte Carlo method. \Vith the Monte Carlo
dramatic changes in its frequency spectmm and timing properties. In method, system of differentird equations that model the circuit are nu-
contrast to nonautonomous/forced systems such as mixers and amplif- mencdly integratd directly in time domain to generate a number of
iers, oscillators exhibit time instability and spectral dispersion. This sample patk for the stochastic processes that model the circuit vari-
phenomenon, peculiar to oscillators, is known as ptie noise or tim- ables. Thus, an ensemble of sample paths is created. Then, by calc-
ing jifter. A perfect oscillator would have Ioctilzed tones at discrete ulatingvarious expectations over this ensemble, one can evaluate var-
frequencies (i.e., harmonics), but any corrupting noise spreads these ious probabilistic characteristics, including correlation finctions and
perfect tones, resulting in high power levels at neighboring frequen- spectrrd densities. If one can prove that the v~tor of stochastic pro-
cies. cesses satisfies some ergodicity properties, as is usurdly the case in
The linear noise anrdysis techniques we described in Section 4.1 practice, it may be possible to crdculate time averages over a single
and Section 4.2 which can be used for amplifiers. filters and mixers are sample path to evrduate some time-averaged probabilistic characteris-
not direcfly applicable to oscillators. In L~ or L~ noise analysis, it tics which provide adequate information in some applications. This
is assumed that the resultant deviation in the response of the circuit due method is referred to as a Monte Carlo method, because in generat-
to the noise sources is small. This assumption is not vrdid for oscilla- ing the sample paths using numericrd integration, one has to realize or
tors or autonomous circuits in general. Oscillators are hauded with a simulate the noise sources or noise-fike desired signals using a random
nordinear pemrbation analysis that is vtild for autonomous systems, number generator. Even though this method may prove to be useful for
wh]ch is summarized next. very specific problems where one can model the circuit with few dif-
An oscillator without noise, i.e., a perfect oscillator, produces a ferentird equations, it is in gened very inefficient and inaccurate. The
periodic signal x~(t). JVhen there are smfl disturbances, or noise, in efficiency can be increasd by using specirdized techniques (e.g., enve-
au oscillator, the perfectly periodic response x~(t) is modifid to x~(t + lope following implemental on top of harmonic brdance for simulation
a(t)) + z(t), where a(t) is a changing time shift, or phe deviation, of nordinear circuits with high-frequency narrowband signals) for the
in the periodic output of the unperturbed oscillator, z(t) is an additive numerical solution of the differential equations that model the circuit,
component, which we term the orbital deviation, to the phase-shifted it is still not efficient and practical enough for rdlstic performance
oscillator waveform. a(r) and z(t) are such that a(t) will, in general, evaluation.
keep increasing with time even if the noise is always small, and if the The ultimate characterization for a stochastic process is ikfiite-
noise sources are removed, a(t) will setfle to a constant vahse, and not dimensioml probability distributions, i.e., the joint probability density
decay to zero. The orbital deviation z(t), on the other hand, will rdways function for a number of its time samples. Note that, in a nordinear
remain small (within a factor of the noise), and if the noise sources are noise amdysis problem, the response of the circuit is not nec~sarily
removed, z(t) will decay to zero. If the circuit is not autonomous, i.e., Gaussian even when the noise sources are Gaussian. This is one of
it is not an oscillator, the phase deviation a(t) will not increase without the basic reasons why nodinear noise anrdysis is (much) more difi-
bound, and will decay to zero if the disturbances are removed. In tils cult than Enear(ized) noise anrdysis. Given a system of differentird
case, representing the response without a phase deviation x~(t) + z t equations model with noise sources for a system, one can derive par-
and using regular linear noise anrdysis to characterize the deviation z [1t tial differentird quations, known as FoMer-Planck equations, for the
is appropriate. probability density of the circuit response. Solving FoWer-Planck-
The discussion above can be formalized for an oscillator described type equations analytically or numerically for probability densities for
by a system of differential equations, and differential equations for a general nosdinear noise anrdysis problem is out of the question. It
the phase deviation a(t) and the orbital deviation z(t) can be derived. is sometimes possible to obtain useful practical results using FoMer-
Then, given the models of the noise sources in the oscillator as stochas- Planck techniques for problems with very specific nature (e.g. cycle-
tic processes, one is interested in computing a stochastic characteriza- sfipping in PLLs, phase noise in oscillators).
tion of the noisy oscillator output x~(t + a(t)) + z(f), i.e., its spectrum. In derding with the nodinw noise anrdysis problem for ~ sys-
It turns out that, for the range of frequencies that are of practical in- tems, the following observation is quite useful: The anrdog front-end
terest, the dominant contribution comes from the phase deviation term of M systems are genetily designed to be as linear as possible from
.r.$(t+ a(t)). hloreover, the orbital deviation term z(t) can be “cleaned the input to the output to prevent distortion of the information signrd,
up” using a limiter at the output of the oscillator, leaving the phase de- considering that the timing and synchronization sigmds such as the LO
viation term unchanged, wh;ch can ody be modified through the use are part of the system and not inputs to it. The signrd path is designed to
of a phase-locked loop. Hence, noise amdysis for oscillators involves be nearly hnear though periodicrdly time-varying as a result of timing
characterizing the phase deviation a(r) as a stochastic procws, which and synchronization signals. Even though the signfl path is designed
is the timingjit~eritself, and dso the resulting spectrum ofx.(t +a(t)). to be as linear as possible, the low level nosdinearity can still have sig-
The phase deviation a(t) dow not change the total power in the per- nificant impact on the performance and needs to be taken into account
iodicsignal x~(t), but it alters the power density in frequency, i.e., the for large noise-like information signals.

202